Panel Threshold Regression Models with

Endogenous Threshold Variables

Chien-Ho Wang
National Taipei University
Eric S. Lin
National Tsing Hua University
This Version: June 29, 2010

Abstract
This paper extends the panel threshold regression of Hansen (1999) and Kourtellos,
Stengos and Tan (2007) to allow for endogeneity of the threshold variable. We consider
the static linear panels with fixed effect. The modified concentrated two-stage least
square methods that are based on inverse Mills ratio bias correction terms are proposed
to estimate threshold parameters. Our estimators are consistent even though individual
effects have any correlation with regressors. Monte Carlo simulations are performed to
investigate the finite sample properties of the proposed estimators.

JEL Classifications: C13, C33, C51

Keywords: Endogenous threshold variables, Panel data, Threshold models

Assistant Professor, Department of Economics, National Taipei University, 151 University Road, San
Shia, Taipei County, 237, Taiwan, R.O.C. Phone: +886-2-8674-7172, Fax: +886-2-2673-9727, E-mail:
wangchi3@mail.ntpu.edu.tw.

Associate Professor, Department of Economics, National Tsing Hua University, Hsin-Chu, Taiwan,
R.O.C. 30013. Phone: +886-3-574-2729, Fax: +886-3-572-2476, E-mail: slin@mx.nthu.edu.tw.

Introduction

One of the most popular topics in nonlinear regression models is the threshold-type regression
model. Threshold models are attractive because they can allow for more flexible regression
functional forms by splitting data with certain unknown threshold values. Although there
are several previous studies on the statistical inference of threshold models, only a few
papers investigate statistical inference under endogenous threshold values. Tong (1983)
first proposes threshold regression models for time series data. Hansen (1999) extends the
threshold regression to static panel data structure and derives the corresponding asymptotic
theory for threshold parameters and regression slopes. Hansen (2000) further shows the
asymptotic properties of concentrated least square estimators for threshold and coefficient
parameters. Caner and Hansen (2004) turn to consider endogenous regressors in threshold
regression models with exogenous thresholds. They use instrument variable methods to
obtain the consistent coefficient estimators. However, all above-mentioned papers virtually
impose the assumption of the threshold variables being exogenous.
To the best of our knowledge, the first paper to investigate threshold regression with
endogenous threshold variables is Kourtellos, Stengos and Tan (below KST, 2007). They
consider the threshold model with endogenous threshold values under cross section data.
When threshold variables are considered endogenous, previous methods cannot obtain accurate threshold parameters, i.e., splitting data based on inaccurate threshold variables will
result in estimation bias in coefficients. KST (2007) interpret the endogenous threshold variables in terms of the traditional sample selection bias. First, they decompose the error terms
into two parts: one is correlated with the endogenous threshold variable, and the other is not.
Second, they use a two-stage least square procedure proposed by Heckman (1979) to rectify
the endogeneity of the threshold variables. Although KST (2007) use the same technology
as the general sample selection model to correct bias, there is one main difference between
the threshold regression model with endogenous threshold and a typical sample selection
model. For a sample selection model, whether the sample is observed hinges on the first
stage selection function, while the entire sample can be observed in the threshold regression
model with endogenous thresholds. KST (2007) utilize a similar two-step approach in sample
selection models to obtain the bias correction parts.
2

Although KST (2007) propose a bias correction method to estimate threshold regression
models with endogenous regressors, their method is in particular developed for cross-sectional
data. Some recent research regarding the issue of investment and growth convergence often
use threshold panel data models, but threshold variables may be correlated with independent
variables. For instance, Hansen (1999) explores the relationship between investment and
financial constraints. Hansen (1999) uses the lagged one period of ratio of long- term debt
to assets for the threshold variable, but the lagged one period of ratio of long- term debt
may be correlated with one of the regressors, e.g., the lagged one period ratio of cash flow
to assets. This correlation can cause inconsistency in traditional panel threshold regressions
with exogenous threshold variables. Kremer, Bick and Nautz (2008) consider the relationship
between economic Growth and inflation. They adopt the five-year average of the annual
percentage change of the CPI index for threshold variable. Although the five-year average
CPI index is a good variable for modeling the threshold effect, still CPI index is possibly
correlated with the growth rate of monetary supply. Therefore, a panel threshold model
which accounts for potentially endogenous variables should be designed for those empirical
research.
In this paper, we attempt to fill up the gap in the literature on panel threshold models
by allowing for endogenous threshold variables. We will propose the two-stage bias correction methods to estimate parameters of panel threshold model with endogenous threshold
variables across different specifications. Three types of panel threshold models are considered in this article: 1) both (main) panel threshold model and threshold variable equation
without individual specific effects; 2) panel threshold model with individual specific effect
but threshold variable equation without individual specific effect; and 3) both panel threshold model and threshold variable equation with individual specific effects. For the panel
threshold model without any fixed effect, we use a bias-correction method to get the consistent estimator. Regarding the main panel threshold model with individual effect only, we
use a difference transformation to eliminate fixed effect. The consistent estimators can be
obtainede by a bias-correction method after the transformation. When both equations have
individual specific effects, we adopt the least squares dummy variable (LSDV) approach to
get the sample residuals sequence in threshold equation. With the sample residuals, we then
conduct the first-difference transformation to remove the fixed effect from panel threshold
3

equation. The bias correction method again will be used to obtain consistent estimators for
threshold value and coefficient parameters.
The remainder of this paper is organized as below: In Section 2, the panel threshold
models without individual effects are introduced. We will propose one consistent estimation
method to handle panel threshold regression with endogenous threshold variables. In Section
3, we relax the assumptions in Section 2 and allow individual effects in main panel threshold
equation. The consistent bias correction estimators will be developed for panel threshold
model with fixed effect and endogenous regressors. In Section 4, we consider a more general
case of the endogenous panel threshold model, the main and threshold equations will exist
individual effects together. A new consistent bias correction estimation method will be
suggested. Finally, some of the future research will be discussed in the last section.

The Panel Threshold Regression with Endogenous

Threshold Models

We consider the balanced panel data throughout the paper, namely, {yit , xit , qit : 1 i n, 1 t T },
where the subscripts i and t are indexes for individual and time, respectively. Consider the
typical panel threshold model as below:
yit = xit I(qit )1 + xit I(qit > )2 + eit

(1)

qit = zit + uit ,

(2)

where qit is an observed threshold variables, is an unknown threshold parameter, and zit
is a vector of instruments.1
Let us define the indicators with respect to the threshold variable qit :

1 iff uit zit
I(qit ) =
0 iff uit > zit

(3)

and

I(qit > ) =

0 iff uit zit

.
1 iff uit > zit

We shall also make the following assumptions in this paper:

1

Note that zit =

z1it

z2it

contain the set of instrument variables, where z2it = xit .

(4)

Assumption 2.1. {yit , xit , qit , eit } is strictly stationary, ergodic.

Assumption 2.2. E|xit |4 < and E|eit |4 < .
Assumption 2.3. n and T is fixed.
Assumption 2.4. For some fixed number G < and 0 < < 1/2, = 2 1 = n G.
Assumption 2.5. Let ft () denote the density function of qit . Define
D() =

T
X

E(Gxit |qit = )ft ()

t=1

D() is continuous at = o .
Assumption 2.6. Assume D(o ) = D, 0 < D < .
Assumption 2.7. uit |zit N (0, 1)
Assumption 2.8. The joint distribution between eit and uit is defined as:
 



 2
eit
0
e j
, j = 1, 2,
|xit , zit N
,
j 1
uit
0

(5)

where j is the covariance between eit and uit ; j = 1 when qit and j = 2 when qit > .

Assumptions (2.1)-(2.3) are standard for fixed-effect panel data regressions. Assumptions
(2.4)-(2.6) are similar assumptions that are used in Hansen (1999). The main purpose of
these assumptions is to simplify the analysis and to get rid of the nuisance parameters in
the distribution of threshold estimations.
Assumptions (2.7) and (2.8) impose the correlation relationship between threshold variable and panel threshold errors. Assumption (2.7) is the same as KST (2007). We are able
to find the accurate functional form of the generated regressor for a two-stage bias correction
estimator with Assumption (2.7). Assumption (2.8) describes the endogeneity structure of
the panel threshold model. From Assumption (2.8) and KST (2007), it is useful to decompose
eit into two parts:

it
uit


=

1 j
0 1
5



eit
uit


.

(6)

We then get the joint distribution of it and uit :

 



 2
e j2 0
it
0
.
|xit , zit N
,
uit
0
0
1

(7)

Under Assumptions (2.1)-(2.8), the effect introduced by endogenous threshold variables,

j uit enters Equation (1) linearly.2 When qit , the conditional expectation of panel
threshold model is:
E [yit |xit , zit , qit ] = E [yit |xit , uit zit ] = xit 1 + 1 ( zit ).

(8)

When qit > , we have that:

E [yit |xit , zit , qit > ] = E [yit |xit , uit > zit ] = xit 1 + 2 ( zit ),

(9)

(zit )
(zit )
where 1 ( zit ) = (z
and 2 ( zit ) = 1(z
. (.) and (.) denote the
it )
it )

density and cumulated density function of a standard normal distribution. 1 ( zit ) and
2 ( zit ) are the well-known inverse Mills, which play the role of bias correction terms. We
can rewrite panel threshold model with endogenous threshold variables for a single equation
form:
yit = xit I(qit )1 + xit I(qit > )2 + (qit , zit , , ) + it ,

(10)

where
(qit , zit , , ) = 1 1 ( zit )I(qit ) + 2 2 ( zit )I(qit > ).
When 1 = 0 and 2 = 0, the bias correction items will disappear. We get Hansen (1999)
panel threshold regression for exogenous threshold models.
yit = xit I(qit )1 + xit I(qit > )2 + it

(11)

Although Equation (10) looks like the error interdependence between main equation and
sample selection equation, there is one important difference between panel threshold models
and sample selection models. In panel threshold model with endogenous models, we can
observe all variables, but in sample selection models, we only observe part of the variables.
2

To simplify our analysis, we consider that the correlation between eit and uit is fixed across i and t, i.e.,
j , in this paper. This setting can be relaxed in the future study.

Our estimation procedure proceeds in three steps: First, we estimate the parameter
in Equation (2) by ordinary least square method.

Second, we estimate the threshold

parameter by minimizing a concentrated least square criterion using

from first stage.
S CLS (i (), i (), )
= arg min
[,]

N X
T
X

(yit xit I(qit )1 xit I(qit > )2 (qit , zit , ,

))2

(12)

i=1 t=1

Third, we estimate the least square coefficient parameters 1 and 2 based on the split

samples implied by .
Corollary 1. For concentrated two-stage least square estimator in the case of endogenous
p
panel threshold regression defined as Equation (12), we have CLS .

The Panel Threshold Regression with Endogenous

Threshold and Fixed Effect

Now, consider the panel threshold model with time-invariant fixed effect as below:
yit = xit I(qit )1 + xit I(qit > )2 + i + eit

(13)

qit = zit + uit ,

(14)

where qit is an observed threshold variables, is an unknown threshold parameter, zit is

a vector. We allow for individual effects i to be arbitrarily correlated with independent
variables xit in Equation (13). When panel threshold model exists time-invariant individual
effect, the individual effect i must be eliminated in advance. For obtaining the consistent
coefficient estimators, individual effects must cancel out before we use two-stage least square
methods. We use the same idea that corrects for attrition bias in unbalanced panel data
from Wooldridge (2002). We use first difference transformation to eliminate fixed effect i :
yit = (xit I(qit )xit1 I(qit1 ))1 +(xit I(qit > )xit1 I(qit1 > ))2 +eit , (15)
3
Because the threshold parameters is crucial for sample splitting, we will first obtain the estimated
threshold parameter . If we want to obtain coefficient estimators with threshold parameter together by
nonlinear regression, it will take a time-consuming calculation. For simplifying our analysis, we will use
two-step estimation procedure suggested by Hansen (2000) and Kourtellos, Stengos and Tan (2007).

where yit = yit yit1 and eit = eit eit1 . To obtain the bias correction estimators, we
need to reset some original assumptions:
Assumption 3.1. {yit , xit , qit , eit } is strictly stationary, ergodic.
Assumption 3.2. The joint distribution between eit and uit is defined as:
 



 2
eit
0
e j
,
|xit , zit N
,
j 1
uit
0

(16)

where j is covariance between eit and uit , j = 1 when qit and j = 2 when qit > .

Assumption 3.3. : E(x0it xit ) has full rank.

Compared with the assumptions in panel threshold models without individual effect, we
can find that eit and uit must have joint normal assumptions, not eit and uit . Under
Assumptions (2.2)-(2.7) and (3.1)- (3.3) satisfied, we can decompose the residuals sequence
eit of transformed equation into two parts:


 

eit
1 j
it
,

uit
0 1
uit

(17)

and


it
uit


|xit , zit N

0
,
0

e2 j2 0
0
1

 
(18)

As the case of panel threshold model without individual effect, panel threshold model with
individual effect can be written as a single equation form:

yit = (xit I(qit ) xit1 I(qit1 ))1 + (xit I(qit > ) xit1 I(qit1 > ))2

(19)

+ (qit , zit , , ) + it ,

where
(qit , zit , , ) = 1 1 ( zit )I(qit ) + 2 2 ( zit )I(qit > )

(20)

When we neglect sample selection parts in panel threshold model, the threshold value
can not be estimated consistently. The ordinary least square estimators for first difference

transformations of threshold panel data are generally inconsistent. Only at 1 = 2 = 0, the

OLS estimators are consistent.
For getting consistent estimators, our estimation procedure proceeds in four steps: First,
we execute first difference transformations for main equation to eliminate individual effect.
Second, we use ordinary least square method to estimate in Equation (2). Third, we
estimate the threshold parameter by minimizing a concentrated least square criterion with

from second stage.

S CLSF E (i (), i (), )
= arg min
[,]

n X
T
X

(yit xit I(qit )1 xit I(qit > )2 (qit , zit , , ))2

(21)

i=1 t=1

Fourth, when is acquired by concentrated least square, we use to split data set and use
least square method to estimate coefficient parameters.
Theorem 1. When Assumption (2.2)-(2.7) and assumption (3.1)-(3.3) are satisfied, for
concentrated two-stage least square estimator in the case of endogenous panel threshold model
p
with fixed effect defined as Equation (21), we have CLSF E .

The Panel Threshold Regression with Endogenous

Threshold and Fixed Effects in Main and Threshold
Equations

In this section we allow both main and threshold equations with fixed effects. The panel
threshold model we consider is
yit = xit I(qit )1 + xit I(qit > )2 + ci + eit

(22)

qit = zit + i + uit ,

(23)

where qit is an observed threshold variables, is an unknown threshold parameter, and zit
is an exogenous variable. Because the fixed effects exist in main and threshold equations,
we cannot obtain consistent estimators 1 and 2 by first difference transformation for two
equations. We propose a modified two stage least square method. First, we use least square
dummy variable regression for threshold equation (23) to obtain
and i . Second, the first
9

difference transformation is used to remove individual effect for main equation. After the
individual effect has been removed in main equation, we can rewrite Equation (22) and (23)
as one equation.
yit = (xit I(qit ) xit1 I(qit1 ))1 + (xit I(qit > ) xit1 I(qit1 > ))2
+(qit , zit , ) + it ,

(24)

where
(qit , zit , , , i ) = 1 1 ( zit i )I(qit ) + 2 2 ( zit i )I(qit > )

(25)

Because the coefficients of bias-correction equation (24) depend on threshold value , the
consistent estimator of can be derived by solving the concentrated least square equation.
S

CLSDF E

(i (), i (), ) = arg min

[,]

n X
T
X

{yit (xit I(qit ) xit1 I(qit1 ))1

i=1 t=1

(xit I(qit > ) xit1 I(qit1 > ))2 (qit , zit , i , ,

)}2

(26)

Equation (22) can be

When we obtain the consistent estimator of threshold parameter ,
estimated by ordinary least square method.
Corollary 2. When Assumption (2.2)-(2.7) and (3.1)-(3.3) are satisfied, the concentrated
two-stage least square estimator in the case of endogenous panel threshold model with fixed
p
effect defined as Equation (26), we have CLSF E .

Monte Carlo Simulation

We conduct a Monte Carlo simulation to study the finite-sample properties of the proposed
threshold panel data estimators against other threshold panel estimators in the literature.
We compare our two-stage panel threshold estimators with the fixed effect panel threshold
estimator by Hansen (1999).The data generation process considered in this paper is based
on following panel threshold regression:

1 xit + ci + eit qit
yit =
,
2 xit + ci + eit qit >
10

(27)

where the threshold value is set to 2. For threshold equation, we consider two cases. The
first one is the threshold equation without fixed effect i :
qit = 2 + 3z1it + 3z2it + uit

(28)

Second, we allow the threshold equation with fixed effect i :

qit = 2 + 3z1it + 3z2it + i + uit ,

(29)

where xit , z1it , and z2it are identical and mutually independent standard normal distributed.
The fixed effects ci and i follow i.i.d. N (0, 1). For controlling the endogeneity, we use the
similar method as in KST (2007) to generate (it , uit ) i.i.d. N (0, I), and eit = e2 (o it +
p
(1 o )uit )/ o2 + (1 o )2 . So that the correlation between eit and threshold variable qit
p
is given by = o / o2 + (1 o )2 . We set o = (0.05, 0.5, 0.9). Regarding the coefficients
of threshold regression, we fix 2 = 1, and vary 1 . We set = 1 2 and examine
= (0.05, 0.5). Finally, we consider n = 50, 100, 200, 500 and T = 10, 20. The number of
replications is 1, 000 for all cases.
Tables 1-2 show the mean square errors (MSE) of the Monte Carlo simulations of the
threshold coefficients and the difference of slope coefficients = 2 1 of two different
estimators. First we consider the estimation of threshold value . No matter the and
are, our panel threshold estimators perform much better than Hansens threshold estimators
under large N and small T in terms of MSE. We also find that panel threshold estimators
using fixed effect transformations have serious downward biases. Second, Tables 3-4 indicate
that the estimators of difference of slope coefficients estimated by fixed effect transformations
are seriously upward biased. Our bias-correction coefficient estimators are consistent under
different and when N goes to infinity and T fixed. In term of thresholds and slope
coefficients our bias-correction estimators ourperform Hansen (1999) fixed effect estimators.

Conclusion

In this paper we extends the panel threshold regression of Hansen (1999) to allow endogeneity
of the threshold variable. We develop a bias-correction two stage least square estimator to
eliminate the endogeneity in the threshold variable. From the Monte Carlo simulations we
11

examine the finite sample performance of our bias- correction estimators with Hansens fixed
effect estimators. Our estimators perform fairly well for a variety of parameters and (N, T )
combinations.

References
[1] Caner, M. and B. Hansen, (2004), Instrument variable Estimation of a Threshold Model,
Econometric Theory, 20, 813-843.
[2] Hansen, B. (1999), Threshold Effects in Non-dynamic Panels: Estimation, Testing, and
Inference, Journal of Econometrics, 93, 345-368.
[3] Hansen, B. (2000), Sample Splitting and Threshold Estimation, Econometrica, 68, 575604.
[4] Heckman, J. (1979), Sample Selection Bias as a specification Error, Econometrica, 68,
575-603.
[5] Kourtellos, A., T. Stengos and C.M. Tan (2007), Threshold Regression with Endogenous
Threshold Variables, Working paper, Department of Economics, University of Cyprus.
[6] Kremer, S., A. Bick and D. Nautz (2008), Inflation and Growth: New Evidence from a
Panel Threshold Analysis, Working Paper, Goethe University Frankfurt.
[7] Neyman, J. and E. Scott (1948), Consistent Estimates Based on Partially Consistent
Observations, Econometrica, 16, 1-32.
[8] Tong, H. (1983), Threshold Models in Non-linear Time series analysis, Lecture Notes in
Statistics No.21, Heidelberg: Springer-Verlag press.
[9] Wooldridge, J. (2002), Econometric Analysis of Cross Section and Panel Data , Cambridge, Massachusetts: MIT press.

12

Appendix: Mathematical Proof

Proof of Theorem 1:

yi2
xi2
..
..
Let yi = . and xi = .
yiT
xiT

y1
..

individuals, Yt = yi and Xt
.

..

yT

. Then let Yt , Xt denote the data stacked over all

x1
..
.

xi . We rewrite Equation (19) as a vector form.

..
.
xT

Y = X2 + X + 2 () + + e,

(30)

where Y = Yt Yt1 , X = Xt Xt1 , X = Xt I(qit ) Xt1 I(qit1 ), () =

I(qit )1 () + (1 I(qit ))2 (), = ()I(qit ) and = 1 2 . When

= (X , ),
= 0 , 0 = (0), 0 = 0 and X0 = X0 . We define X()
= (X, ()), X
= (X
, (X
X
)) like KST (2007).
X

.
We define the projection matrix P and orthogonal matrix M

(X
0 X
)1 X
0
P = X

(31)

= I P
M

(32)

Equation (30) can be rewritten as:

0 + e
Y = X(0)
+X




2
2
under = o , where =
and =

Then,

(33)

n1+2 T 1 (S CLSF E () e0 e)
Y e0 e)
= n1+2 T 1 (Y 0 M
=n

1+2

0 + e)0 M
(X(0)

0 + e) e0 e
(X(0)
+X
+X
13

X
0 )C2
X
)1 (X

)(X

)(X
X
)1 (X
X
0 )C1 + C2 (X(0)
X
X
= (nT )1 [C1 (X(0)
0
0X
0 1 0
+C3 (X
0 )(X X ) (X X0 )C3 ] + op (1)

We can derive the following result:

0
0 X
)(X
0 X
)1 (X
0 X
)C] + op (1).
n1+2 T 1 (S CLSF E () e01 [C (X
0

When [o , ],

0 X
0
0 X

X
X
p
0

1 X X0

)(
) (
)C] b1 ()
[C (
nT
nT
nT
0

Under Assumptions 2.2-2.7 and 3.1-3.3 are satisfied,

d
b ()
d 1

(34)

(35)

> 0. Because b1 () > 0 is

monotonic increasing and continuous at 0 , b1 () > 0 will have uniquely minimized at 0 .

Similarly, When [, 0 ],
0 X
X
0 X

0
X
p
0

1 X X0
[C (
)(
) (
)C] b2 (),
nT
nT
nT
0

(36)

b2 () > 0 will have uniquely minimized at 0 .

Under [, ],
p

n1+2 T 1 (S CLSF E () e0 e) b1 ()I( > 0 ) + b2 ()I( 0 ),

(37)

where 0 < < 1/2. We can obtain

p
CLSF E .

Proof of Corollary 2:
Corollary 2 is a special case of Theorem 1. Please refer to the proof for Theorem 1.

14

(38)

Table 1: Finite-sample Performance for Threshold Value Estimation (1 2 = 0.05)

Method
= 0.05
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE
Method

= 0.5
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE
Method

= 0.9
N = 50
N = 100
N = 200
N = 500

T = 10

T = 20

mean median
MSE
mean median
MSE
-0.665
-0.036 24961.78 -0.628
0.265 25832.25
-11.345
-11.16 181688.4 -12.43 -12.257 210775.7
-0.546
0.282 24626.43 -0.289
0.732 22597.33
-12.382 -12.185 209352.7 -13.285 -13.066 236098.5
-0.265
0.968 23640.77
0.622
1.508 14370.53
-13.232 -13.232 234145.7 -14.115
-13.92
261835
0.804
1.726 13200.37
1.57
1.961 5116.832
-14.396 -14.209
270869 -15.181 -15.034 297085.8
T = 10

T = 20

mean median
MSE
mean median
MSE
1.76
1.843 4180.392
1.868
1.855 2629.118
-11.388 -11.183 181967.1 -12.502 -12.336 212850.7
1.984
1.982 2716.145
1.986
2.057 1914.692
-12.347 -12.188
208403 -13.316 -13.139 236969.3
2.096
2.055 2163.062
1.994
2.013 1368.487
-13.213 -12.977 233551.9 -14.142 -13.948 26267.85
1.892
1.946 1272.687
2.017
2.018 831.415
-14.432 -14.236 272146.5 -15.204 -15.011 297784.2
T = 10

T = 20

mean median
MSE
mean median
MSE
pthet
1.968
1.952 2903.662
1.94
1.97 1950.399
FE -11.404 -11.191
182333 -12.457 -12.226 211744.9
pthet
2.087
2.061 2039.976
1.987
1.992 1346.099
FE -12.407 -12.201 210181.9 -13.177 -13.003 132646.9
pthet
2.053
2.07 1388.162
1.98
1.991 810.065
FE -13.233 -13.022 233997.3 -14.21 -14.029 264870.4
pthet
2.032
2.01 768.197
2.015
2.034 488.923
FE -14.43 -14.213 272078.7 -15.15 -14.985 296049.4

15

Table 2: Finite-sample Performance for Threshold Value Estimation (1 2 = 0.5)

Method
= 0.05
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE
Method

= 0.5
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE
Method

= 0.9
N = 50
N = 100
N = 200
N = 500

T = 10

T = 20

mean median
MSE
mean median
MSE
1.981
1.986 166.392
2.001
1.996
22.96
-11.51 -11.381 185594.5 -12.452
-12.28 211379.6
2.001
1.997
26.769
1.996
1.996
4.78
-12.348 -12.196 208522.5 -13.35 -13.218 237957.2
1.997
1.998
5.771
1.998
1.998
1.294
-13.246 -13.081 234529.1 -14.171 -13.946 263563.2
1.998
1.999
0.954
1.999
1.999
0.225
-14.31 -14.088 268120.9 -15.23 -15.059 298604.3
T = 10

T = 20

mean median
MSE
1.956
1.974 588.074
-11.397 -11.204 182358.2
1.987
1.991 147.534
-12.403
-12.22 209992.5
1.994
1.995
39.786
-13.218 -13.032 233865.7
1.997
1.999
6.236
-14.353 -14.207 269418.1

mean median
MSE
2.01
2.005 146.177
-12.519 -12.303 213241.6
2
1.996
37.577
-13.288 -13.061 235957.1
1.997
1.999
9.798
-14.061 -13.888 259966.5
2
1.999
1.536
-15.198 -14.978 297707.8

T = 10

T = 20

mean median
MSE
mean median
MSE
pthet
2.007
1.978 586.007
2.017
2.002 152.444
FE -11.449 -11.245 183448.1 -12.462 -12.259 211568.1
pthet
2.005
2.001 163.956
2.002
1.994
35.211
FE -12.281 -12.117 206305.2 -13.229 -13.084 236487.9
pthet
1.997
1.999
46.364
1.997
1.998
1.177
FE -13.332
-13.1
237440 -14.134 -13.936 262362.6
pthet
1.997
1.999
6.466
2
1.999
0.147
FE -14.31 -14.072 268027.8 -15.18 -14.979 297017.2

16

Table 3: Finite-sample Performance for Main Regression Coefficient (1 2 = 0.05)

Method
= 0.05
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE

T = 10
mean
0.066
1.026
0.038
1.025
0.006
1.026
0.061
1.024

Method
= 0.5
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE

N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE

median
MSE mean median
MSE
0.072 210.707 0.062
0.059 296.961
1.029 955.946 1.026
1.026
956.9
0.064 363.922 0.046
0.059 745.058
1.023 953.157 1.025
1.025 952.595
0.06 6352.047 0.045
0.054 48.168
1.026 953.414 1.025
1.025 951.868
0.054
89.671
0.05
0.051
0.66
1.024 949.718 1.024
1.023 949.313
T = 10

mean
0.037
1.025
0.043
1.023
0.045
1.025
0.045
1.024

Method
= 0.9
N = 50

T = 20

mean
0.046
1.025
0.045
1.024
0.045
1.025
0.046
1.025

median
0.039
1.024
0.043
1.023
0.047
1.026
0.044
1.025

T = 20
MSE mean
6.476 0.044
953.712 1.024
2.752 0.044
948.686 1.024
1.326 0.044
951.615 1.025
0.518 0.046
949.627 1.025

median
0.045
1.022
0.045
1.024
0.044
1.024
0.047
1.025

T = 10

T = 20

median
0.045
1.025
0.043
1.025
0.043
1.024
0.045
1.026

MSE mean median

5.483 0.044
0.044
953.695 1.027
1.028
2.51 0.043
0.042
949.796 1.024
1.025
1.291 0.045
0.044
951.191 1.026
1.028
0.486 0.046
0.046
951.089 1.025
1.025

17

MSE
2.826
951.684
1.167
951.048
0.621
950.783
0.257
951.217

MSE
2.309
959.098
1.236
951.172
0.553
953.335
0.234
950.986

Table 4: Finite-sample Performance for Main Regression Coefficient (1 2 = 0.5)

Method
= 0.05
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE

T = 10
mean
0.507
1.25
0.504
1.25
0.502
1.25
0.5
1.25

Method
= 0.5
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE

mean
0.485
1.251
0.492
1.25
0.498
1.25
0.498
1.25

Method
= 0.9
N = 50
N = 100
N = 200
N = 500

pthet
FE
pthet
FE
pthet
FE
pthet
FE

mean
0.481
1.249
0.488
1.247
0.495
1.251
0.497
1.25

median
0.511
1.251
0.504
1.249
0.502
1.25
0.5
1.25

T = 20
MSE mean
4.431 0.504
567.927
1.25
2.154 0.502
565.006 1.248
1.067 0.501
563.493 1.249
0.447 0.499
562.583 1.251

median
MSE
0.506
1.961
1.25 566.405
0.502
1.085
1.248 561.43
0.501
0.528
1.25 562.564
0.499
0.214
1.251 564.662

T = 10

T = 20

median
MSE mean
0.489
6.411 0.489
1.252 568.459 1.249
0.495
2.632 0.495
1.25 564.915 1.251
0.498
1.219 0.498
1.25 564.07
1.25
0.498
0.49 0.499
1.249 562.307 1.249

median
0.49
1.249
0.496
1.252
0.498
1.25
0.499
1.249

T = 10

T = 20

MSE
2.69
564.645
1.273
566.119
0.554
563.196
0.213
561.724

median
MSE mean median
MSE
0.483
6.197 0.487
0.487
3.039
1.25 565.868 1.252
1.25 569.449
0.488
2.978 0.494
0.494
1.221
1.247 560.682 1.252
1.251 567.521
0.495
1.374 0.498
0.498
0.536
1.25 564.748 1.251
1.25 565.302
0.497
0.487 0.499
0.5
0.215
1.25 563.459
1.25
1.251 563.469

18